Attribution models

Attribution Readout — 2026-09-11

Pillar: (A) Interpretation. You are choosing a model and reconciling three reports; you are not building first-party tracking. Nothing below needs engineering.

Business type: Ecommerce / DTC, short self-serve cycle, Shopify checkout, spend concentrated in Meta + Google. The DTC playbook applies throughout.

What I do not have: the actual numbers. You have not shared the Shopify order count, the Google Ads conversion count, the Meta conversion count, the date range, the attribution windows, or the spend. Every cell below that needs one of those is left as "—". The structure and the method are complete; you fill the cells, and the readout becomes the decision document.

The question

Should budget move between Google Ads and Meta, and in which direction?

Underneath it, the real question: how many of the orders each platform claims did it actually cause? Right now you cannot answer that, because both platforms are counting the same orders, and the sum exceeds the store total. That is not a bug in either dashboard. It is what platform attribution always does. The fix is not to make the three numbers match. They never will. The fix is to anchor on one number, de-dupe the other two against it, and then run one test that settles the overlap.

Do not reallocate on the current dashboards. Moving spend toward whichever platform reports the higher conversion count rewards whichever platform over-claims harder.

Source of truth

Shopify defines the conversion count. It is the system where the money is real: an order either exists there or it does not. You have no CRM, and for DTC you do not need one for this; the store is the backend.

Consequences, stated bluntly:

• The Shopify order total for the period is the ceiling. Paid orders cannot exceed it; paid orders plus organic, email, direct and everything else must equal it.
• Google Ads and Meta explain *where some of those orders came from*. They do not get to redefine *how many orders there were*.
• Never sum Google and Meta. If Google claims an order and Meta claims the same order, you have one order with two claimants. That is the whole reason your combined platform count is above the store total.

Two Shopify things to check before you fill in the table:

1. Which Shopify number are you using? Total orders includes orders from every channel plus manual/draft orders and possibly POS. For this exercise use online-store orders for the exact same date range the ad platforms report on, and decide once whether refunds/cancellations are in or out. Use the same rule on all three reports.
2. Shopify's own marketing attribution (Marketing / Sales attributed to marketing reports) uses its own model and window. It is a fourth opinion, useful as a last-non-direct-click view, but it is not the source of truth either. The order count is; Shopify's channel split is one more input.

What each source says

De-duped against the Shopify total, never summed. Columns adapted from the skill's template: there is no GA and no CRM here, so Shopify stands in for the CRM column, and Self-reported is empty until you install a survey.

How to fill it, in order:

Step 1. Align the counting rules before comparing anything. The three reports are currently measuring different things. Line them up:

Until these match, the numbers are not even comparable, let alone reconcilable. The date-of-click vs date-of-order difference alone puts orders in different weeks across the three reports. Verify each window in the actual account settings rather than assuming a default; defaults change and accounts get edited.

Step 2. Strip what Shopify structurally cannot see. View-through conversions (Meta 1-day view, Google view-through/engaged-view) are conversions where the person never clicked. Shopify will never attribute those to the platform because there is no click to land on. Pull them out into their own row so the click-based numbers can be compared like with like. Keep the view-through counts visible; they are not fake, they are simply unverifiable from your side.

Step 3. De-dupe the click conversions against Shopify. With windows aligned and view-through separated, compare each platform's click conversions to Shopify's attributed orders for that channel. The remaining excess is the platform claiming orders that either (a) another channel closed after the click, (b) landed outside Shopify's window, (c) were modeled/estimated post-iOS rather than observed, or (d) were counted twice by the platform itself (duplicate purchase events from pixel + CAPI without deduplication is a common cause; that is an ads-skill fix, not an attribution one, but check it, because it inflates the platform count before any attribution logic runs).

Step 4. Report the gap, not a forced tie-out. The output sentence is: *"Google and Meta together claim — purchases. Shopify records — online orders in the same window, of which — are attributable to any paid click. The delta of — is view-through + over-claiming + modeled conversions + window mismatch. Here is how we allocate the overlap."*

Model comparison

The skill's side-by-side (first-touch vs last-touch) is built for long cycles. Yours is short, so the comparison is smaller, but it still exists and still carries the insight.

Recommended model: last non-direct click, applied consistently in one place, with Shopify as the order count.

Why this and not the alternatives:

• Last-touch / last non-direct click is defensible for quick-turn DTC. The closing click really is most of the story for an impulse or low-consideration purchase. It gets less defensible as AOV rises and consideration lengthens, because it over-credits retargeting and branded search. If this brand sells a considered, higher-AOV product, note that and lean harder on the survey and the test below.
• Data-driven (Google's default) is a black box that only sees Google's own touches and needs volume. It cannot see Meta, so it cannot arbitrate Google vs Meta, which is exactly the decision you are trying to make. Leave it on for in-platform bidding if you like; do not use it for cross-channel budget truth.
• First-touch, linear, time-decay, position-based add little on a journey that is often one or two touches long. The information they would carry (what created the demand) is better captured by the survey.

What to look at side by side: each platform's own claimed count against Shopify's last-non-direct-click count for that same platform. That gap is the model comparison for a DTC brand. A platform whose own number sits far above its Shopify-attributed number is either heavy on view-through and modeled conversions, or is taking credit for orders another channel closed. Meta prospecting and retargeting usually show the wider gap; Google Shopping and non-brand search usually the narrower. Confirm with your numbers rather than assuming.

One caveat to name now: if a large share of Shopify orders land as direct or branded search, that is not a channel. It is demand created somewhere else (often Meta, often organic/social/word of mouth) cashing out. Last-click will hand that credit to branded search and direct. Segment branded from non-branded Google spend before you compare Google to Meta, or Google will look better than it is.

Confidence & gaps

• The attribution gap is expected and permanent. Two platforms claiming more than the store's total is the normal state. You can shrink it and explain it; you cannot close it.
• Platform ROAS is fine for the algorithm, not for the budget. Post-iOS ATT both platforms model and estimate conversions and use generous windows. Reported ROAS runs above incremental ROAS. Keep it for in-platform optimization; do not let it decide the Google-vs-Meta split.
• View-through is unverifiable from Shopify. It is not zero, and it is not what the platform says either. Its true value only shows up in a holdout test.
• Self-reported: currently no signal. No post-purchase survey means dark social, organic TikTok/Instagram, word of mouth and podcast-style touches are invisible. Whatever they drive is landing in direct and branded search today.
• AI-search traffic arrives as direct or branded search and is invisible here. Named as a blind spot only; deeper work belongs to ai-seo.
• MMM is not the answer at this stage unless blended monthly spend is material (the reference puts the threshold around six figures a month) and you have two to three years of weekly data with real spend variation. Below that, good UTMs plus a survey plus one test does more.
• Confidence in a reallocation right now: low. The inputs disagree, the counting rules are not yet aligned, and there is no out-of-model check. Confidence rises to medium after Steps 1 to 4 above, and to high only after the incrementality test.

Recommendation

1. This week: align and de-dupe (Steps 1 to 4). Same date range, purchase-only conversion actions, view-through separated, click conversions compared to Shopify's attributed orders per platform. Fill the table. Pull the platform numbers into one sheet (Supermetrics or Coupler if you want it repeatable; a manual export is fine once). Confirm pixel + CAPI deduplication on Meta while you are in there.

2. Adopt one model and stop switching. Last non-direct click, read in Shopify, is the operating model. Platform dashboards stay open for in-platform optimization only. State this to the client in writing, including the sentence "every model is an opinion; this is the one we will be consistent with."

3. Install a one-question post-purchase survey on the Shopify order-confirmation page ("How did you first hear about us?", pick-list of your known channels plus a free-text "other"). Ask at the moment of purchase. This is the out-of-model check that tells you what is hiding in direct and branded search, and the cheapest tiebreaker you can get. Treat it as triangulation, not gospel; people credit the memorable touch.

4. The tiebreaker test worth running: an incrementality test on the channel with the widest platform-vs-Shopify gap. Two platforms fighting over the same orders is precisely the case the skill reserves incrementality for. Options, cheapest first:

• On/off pulse test: pause the disputed campaign type (retargeting is the usual candidate, then branded search) for a defined window and watch total Shopify orders, not platform conversions. Crude, but it directly answers "would these orders have happened anyway?"
• Geo holdout / geo-lift (Meta GeoLift, Google geo experiments) on the bigger of the two platforms if volume allows a clean split.
• Read the result as lift = exposed rate minus control rate, check the interval before believing the point estimate, and translate to incremental CPA = spend ÷ incremental orders. That number, not reported CPA, is what moves the budget. Expect it to be worse than either dashboard says; that is the point.

5. Reallocate only after step 4, and move budget in steps rather than a single swing. Re-read the de-duped table after each move and watch for directional agreement across the three sources; if Shopify, the survey and the platforms all say the same channel is trending up, trust the trend even though the absolute numbers will never match.

What I need from you to turn this into the actual allocation call: Shopify online-order count for the period; Google Ads purchase conversions with the window and model noted, view-through shown separately; Meta purchase conversions with the attribution setting noted, 1-day view shown separately; spend per platform; AOV; and whether branded search is split out. With those, the "—" cells fill in and the Recommendation section gets numbers and confidence levels attached.